diff options
author | Todd Gamblin <tgamblin@llnl.gov> | 2020-11-02 13:21:11 -0800 |
---|---|---|
committer | GitHub <noreply@github.com> | 2020-11-02 13:21:11 -0800 |
commit | ecc3bfd48471035ec260c37e82a045e079addb64 (patch) | |
tree | 133d577ef8652285d6ae3920145177491860323f /.coveragerc | |
parent | 2ca894bd1ea847d7226c7b4479d26cd645071f4e (diff) | |
download | spack-ecc3bfd48471035ec260c37e82a045e079addb64.tar.gz spack-ecc3bfd48471035ec260c37e82a045e079addb64.tar.bz2 spack-ecc3bfd48471035ec260c37e82a045e079addb64.tar.xz spack-ecc3bfd48471035ec260c37e82a045e079addb64.zip |
Bugfix - hashing: don't recompute full_hash or build_hash (#19672)
There was an error introduced in #19209 where `full_hash()` and
`build_hash()` are called on older specs that we've read in from the DB;
older specs may not be able to compute these hashes (e.g. if they have
removed patches used in computing the full_hash).
When serializing a Spec, we want to generate the full/build hash when
possible, but we need a mechanism to skip it for Specs that have
themselves been read from YAML (and may not support this).
To get around this ambiguity and to fix the issue, we:
- Add an attribute to the spec called `_hashes_final`, that is `True`
if we can't lazily compute `build_hash` and `full_hash`.
- Set `_hashes_final` to `False` for new specs (i.e., lazily
computing hashes is ok)
- Set `_hashes_final` to `True` for concrete specs read in via
`from_node_dict`, as it may be too late to recompute hashes.
- Compute and write out all hashes in `node_dict_with_hashes` *if
possible*.
Effectively what this means is that we can round-trip specs that are
missing `_build_hash` and `_full_hash` without recomputing them, but for
all new specs, we'll compute them and store them. So Spack should work
fine with old DBs now.
Diffstat (limited to '.coveragerc')
0 files changed, 0 insertions, 0 deletions